Generating realistic perturbed hydrometeorological time series to inform scenario-neutral climate impact assessments

Scenario-neutral approaches are used increasingly as a means of stress-testing climate-sensitive systems to a range of plausible future climate conditions. To ensure that these stress-tests are able to explore system vulnerability, it is necessary to generate hydrometeorological time series that represent all aspects of plausible future change (e.g. averages, seasonality, extremes). A promising approach to generating these time series is by inverting the stochastic weather generation problem to obtain weather time series that capture all the relevant statistical features of plausible future change. The objective of this paper is to formalize this inverse approach to weather generation, by both characterizing the process of optimizing weather generator parameters and proposing a numerically efficient solution that exploits prior knowledge and accounts for the complexity of the optimization landscape. The proposed approach also provides a structured way to ensure the physical realism of the generated weather time series, by using penalty-based objective functions to focus the optimization on the climate features deemed most relevant to the system being analyzed. A case study in Adelaide, Australia, is used to demonstrate specific implementations of this approach. The use of bounds on the weather generators dramatically decreases the time taken to create time series, and the use of penalties is shown to allow for change in some statistics to be prioritized, while still ensuring the realism of the time series.